Universal Modder's new MIT toolkit NPCTalk lets a coding agent wire talking NPCs into a single-player game: you speak, the NPC answers in character in a cloned Fish Audio Drama 3 voice about 1.2–2.2 seconds later, and acts on what you said. GTA V example, latency budget, costs and voice scores inside.
USC-led 3DCodeVerse open-sources an Apache 2.0 harness that makes Claude Code, Codex or Gemini agents write Blender, Three.js, URDF and GLSL assets, then runs, measures and judges them. Costs, commands, and what the gated dataset holds today.
fal launches Worlds plus realtime WebRTC endpoints for Lingbot World 2 and Abot World 0: start from an image, steer with WASD and camera commands, $0.00125 per compute second. The WMA layer is still experimental.
AlayaLab open-weights MeshAxiom-Base: one image plus a face budget becomes a native triangle mesh — vertices and connectivity generated together, not decimated later. Self-host on CUDA; license still pending.
A solo dev's new MIT tool turns a sentence or a picture into a textured, auto-rigged, animated 3D asset — plus pixel sprites — exported to Godot, Unreal or Unity, fully offline on one 16 GB NVIDIA card.
over|yonder Espalier-2 open-sources MIT patches that run Microsoft's TRELLIS.2 image-to-3D at full 1536 cascade on a 16 GB RTX 5070 Ti — Microsoft lists 24 GB as the minimum. Clone, install.sh, get a textured GLB.
Scenario open-sources its Blender 5 extension (GPL-3.0): image, video, 3D, PBR, audio, 360° environments and text-to-greybox Blockout in the viewport, plus a local MCP server agents can drive. Generation runs on Scenario's paid API.
Skywork open-sources Matrix-Game 4.0 (Apache-2.0): executable low-poly games plus streaming neural rendering for Medieval Castle, Ancient Egypt, and City Racing — with single-GPU inference, browser demo, and HF weights.
Tencent ARC Lab and HKU open Multics (Apache-2.0): one checkpoint jointly generates first-person video of a shared scene for 4–10 players from each view, prompt, actions, and camera path. Code and ~30 GB weights are live.
Artificial Analysis launches AA-3D: 21 models work as agents in an offline sandbox to build interactive three.js scenes, including pinball, dodgeball and glider sims, then get ranked on 10K+ blind human votes. Claude Sonnet 5.5 leads at 1194 Elo in a four-way tie. The best open-weights model is 200+ points back.
Microsoft Research open-sources Code Video Model (Apache-2.0): a coding agent writes a Three.js scene, the proxy render steers a frozen MiniMax-H3 for the first 5 of 20 steps, and you get physics-true, camera-exact 960×544 clips. Five runnable examples plus CodeVideoBench.
ZJU3DV releases InfiniSplat v2 code, 6.6 GB weights and an HF demo: 2–50 overlapping photos or a video, posed or unposed, become a 3D Gaussian scene in one forward pass. PLY + HTML export; free research use, project use needs registration.
NTU, Oxford VGG and ETH open Artisan (Apache-2.0): design, coding and review agents build articulated 3D assets, then a physics loop cuts joint collisions to 1.7%. Exports URDF, MJCF, USD and articulated GLB.
MBZUAI researchers open code + two MiniMax-H3 LoRAs that make Code World Model stream causally: 0.71 s chunks from a live code proxy, 4-step DMD at 5.1 s/chunk on one H200.
Tencent and NTU's GameDirector just shipped code and Apache-2.0 weights: a Gemma 4 E2B director keeps HP, cooldowns and boss AI honest while a Wan2.2-based world model renders the fight.
NVIDIA dropped StreamRender-H3 with full Hugging Face runtime weights and Sana code — a Three.js game engine streams semantic maps, a two-step MiniMax-H3 student renders photoreal video in real time.
PKU, Math Magic, and HKUST dropped UniForge3D with Apache-2.0 code and Hugging Face weights — sparse unposed photos in, complete textured 3D assets out, runnable locally today.
UChicago researcher Rana Hanocka's Thrixel came out on Oct 8 with $3M and a 3D generator whose models arrive as named, editable parts, plus a Build World skill that lets Claude Code, Codex or Cursor build a playable game around them.
Two production-minded 3D launches landed on Oct 8: Scenario's Claude Opus 5.5 3D outputs named, separate parts in a real Blender file, and Dreamlands 1.0 opens AI world building to everyone from $20 a month.
Applied Intuition open-sourced Building Rome: TRELLIS.2 LoRA deltas that turn one photo into a full metric scene GLB with occluded geometry filled in. Code and weights out today, CC BY-NC.
Odyssey launched Odyssey-3 on Oct 8: an autoregressive world model that builds an interactive environment from a prompt and reacts in real time as you move or trigger events. A free research preview is live in the browser, and the Pro tier posts a record 66.1 on Physics-IQ Verified.
Charakuru, a free Tripo-sponsored kit from indie developer Nano, turns a single full-body character illustration into a rigged VRM 1.0 avatar: a local review app plus seven agent Skills hand Tripo, Blender and rigging to Codex while you pick candidates. Version 1.0.3 shipped Oct 8 with MToon shading.
HKU MMLab released the MIT-licensed code for Recursive Game Creator on Oct 7: four agents design, build, playtest and review an HTML or Godot game over several rounds, keeping a new version only when it wins blind A/B review, with optional Hunyuan image and 3D assets.
AssetFurnace, the open-source studio formerly called Image to 3D Lab, shipped 0.4.0 on Oct 7: type a sentence or drop a picture and get a textured, auto-rigged character with 21 preset moves as a GLB for Godot, Unity, Unreal or Blender, all on your own Mac or NVIDIA PC.